DeepSeek R1

DeepSeek R1 is a 671-billion parameter Mixture-of-Experts reasoning model released on January 20, 2025, trained with large-scale reinforcement learning to encourage natural emergence of reasoning behaviors like self-verification and reflection. With 37 billion activated parameters and a 64K token context window, it reaches performance on par with OpenAI o1 across mathematics, coding, and reasoning tasks. Benchmark results include 79.8% on AIME 2024 and 97.3% on MATH-500. The model outputs explicit chain-of-thought reasoning in thinking tokens between think tags alongside final answers, and supports tool calling, making it suitable for agentic workflows and complex reasoning applications. Released under the MIT license, DeepSeek R1 is fully open-source and ships with distilled variants from 1.5B through 70B parameters based on Qwen and Llama architectures, broadening access to reasoning-grade capabilities for developers.